# AI-Native Support

AI is the core of the Karmaflow.ai helpdesk, not an add-on. The same AI agents that power your voice, chat, SMS, and email are first-class actors inside support: they read every incoming ticket, propose the next step, draft grounded replies, and — within a policy you define — resolve routine issues on their own. What makes this usable in a real support operation is that every AI action is **explicit, governed, and audited**, with a human always in reach.

This article explains how that works and why it matters. Configuration steps live in the signed-in documentation.

## The Three Principles

Everything below follows from three rules that hold everywhere in the helpdesk:

1. **AI-vs-human is always explicit.** Every message and action records who — or what — produced it. AI replies are labeled as AI in the agent workspace, the customer portal, and outbound email. The AI never poses as a person.
2. **A human is always reachable.** Customers can press **Talk to a human** on any AI-handled ticket and it escalates immediately.
3. **Humans stay above the loop.** AI acts only inside the envelope you configure, and everything it does is written to the audit trail.

## What the AI Does

### Triage on every ticket

The moment a ticket is created, an AI agent analyzes it and produces a structured read: a concise summary, detected sentiment, a suggested priority and category, the specific information still missing from the customer, similar past tickets, and the most relevant knowledge-base articles. These are **suggestions** — they inform the human (or a governed action) but never silently change the ticket.

### Grounded reply drafting

Agents can generate a suggested reply grounded in the conversation and your published knowledge base. Because it's grounded in *your* content, the draft reflects your policies and product — not a generic guess. The draft lands in the composer for a human to review, edit, and send; nothing is sent automatically from the composer.

### Governed autonomous resolution

Where you allow it, an AI agent can act on a ticket directly — reply, request missing information, change status, or resolve it — closing simple, high-confidence issues without a human touch. This is where deflection comes from, and where governance matters most.

## How Governance Works

Every action an AI agent wants to take on a ticket is classified by your policy into one of three outcomes:

```mermaid
flowchart TD
  action[AI proposes an action] --> policy{Your policy}
  policy -->|Allowed| run[Runs immediately<br/>labeled as AI]
  policy -->|Needs approval| queue[Human approval queue]
  queue -->|Approved| run
  queue -->|Rejected| human
  policy -->|Forbidden| human[Escalates to a human]
  run --> audit[(Audit trail)]
  human --> audit
```

| Outcome | What happens |
|---|---|
| **Allowed** | The AI performs the action immediately — still labeled as AI and written to the audit trail |
| **Needs approval** | The action is held in a human approval queue for a person to approve or reject before it runs |
| **Forbidden** | The AI may not take the action; it escalates to a human instead |

You set this per action type, and you tune it with guardrails:

- An **autonomy master switch** that, when off, routes even "allowed" replies and resolutions to the approval queue — a conservative way to start.
- A **confidence threshold**: below it, an otherwise-allowed action is downgraded to "needs approval."
- **Sensitive topics** (e.g. legal, billing disputes, cancellations, security): when triage detects one, AI actions on that ticket always require human approval.

A common starting posture is to let AI add internal notes and request information freely, require approval for public replies and resolutions, and forbid the rest — then loosen as the AI earns trust. Because every action is attributed and logged, you can *see* that trust build in the numbers.

## Grounding and Citation

AI answers are only as good as what they're grounded in. When you publish a knowledge-base article, it's synced into the tenant's AI knowledge store, and from then on your agents can retrieve and **cite** it when answering. Keeping the knowledge base current is the single highest-leverage way to make AI replies accurate — and the reporting surfaces which articles the AI cites most and where content gaps are costing you tickets.

## Why It Matters for Evaluation

The hard part of AI support isn't generating a reply — it's doing so safely, at scale, in a way a support leader can stand behind. Karmaflow.ai's answer is explicit attribution, per-action policy, confidence and sensitivity guardrails, a human approval queue, and a complete audit trail. You get the deflection and speed of automation without giving up control or accountability.

> **Want the configuration details?** The AI governance policy, approval queue, and knowledge-base setup are covered step-by-step in the signed-in documentation. [Sign in to Karmaflow.ai](https://omni.karmaflow.ai/auth/login) to read them.